Large-scale exhibition crowd density monitoring and abnormal behavior identification method and system

Abstract

The invention relates to the technical field of security and protection monitoring, in particular to a large-scale exhibition crowd density monitoring and abnormal behavior recognition method, which comprises the following steps: acquiring real-time video data acquired by a plurality of monitoring devices in an exhibition venue, and performing crowd density estimation on the real-time video data to generate a real-time crowd density distribution map; acquiring historical crowd flow data, acquiring a crowd dynamic evolution trend in a future preset time period according to the real-time crowd density distribution diagram and the historical crowd flow data, and identifying and acquiring a potential bottleneck region based on the crowd dynamic evolution trend; according to the potential bottleneck area and the real-time crowd density distribution diagram, a space attention thermodynamic diagram is generated, real-time mastering of the crowd state in the exhibition venue, risk pre-judgment and accurate abnormity disposal can be achieved, management and control measures in a large exhibition high-density crowd scene are more timely, and the response efficiency is higher.

Description

Technical Field

[0001] This invention relates to the field of security monitoring technology, and in particular to a method and system for monitoring crowd density and identifying abnormal behavior at large-scale exhibitions. Background Technology

[0002] With the continuous development of smart cities and public safety management, crowd safety management for large-scale exhibitions and conventions is receiving increasing attention. In the field of operation and management of large-scale exhibitions and conventions, crowd density monitoring and abnormal behavior identification are key technical aspects to ensure the safe and orderly conduct of the event. Currently, the industry has developed various technical solutions for crowd management. Among them, crowd density monitoring is mostly based on surveillance video data. The mainstream method is to collect real-time images through network cameras deployed in the exhibition venue, and then use image recognition technology to obtain information on the number or density of people in different areas of the venue. Some solutions also combine crowd flow statistics collected by devices such as turnstiles and infrared sensors to help determine the crowd gathering situation in key areas, such as entrances and exits, key passages, and popular exhibition areas.

[0003] Although existing technologies can achieve basic crowd density monitoring and abnormal behavior identification, they still have the following shortcomings when adapting to the complex scenarios of large-scale exhibitions. Existing technologies mainly focus on monitoring individual behavioral characteristics and crowd density distribution. They judge abnormalities by detecting individuals' rapid running or irregular collisions in video footage, or identify potential risks based on the overall macroscopic movement state of crowd gathering. Methods based on individual behavioral characteristics have low accuracy when the crowd density is high and there is severe mutual occlusion. Methods based on crowd density distribution detection focus more on static analysis of the current state. However, crowd movement in large-scale exhibitions has obvious spatiotemporal correlations. Simply analyzing the current state is difficult to predict potential gathering risks, resulting in a lag in control measures. This makes it impossible for security personnel to quickly formulate targeted response strategies, affecting the efficiency of emergency response. Summary of the Invention

[0004] The main objective of this invention is to provide a method and system for monitoring crowd density and identifying abnormal behavior at large-scale exhibitions, aiming to solve the technical problems mentioned in the background section.

[0005] This invention proposes a method for monitoring crowd density and identifying abnormal behavior at large-scale exhibitions, comprising: The system acquires real-time video data collected by multiple monitoring devices within the exhibition venue, and generates a real-time crowd density distribution map by estimating the crowd density of the real-time video data. Historical population flow data is obtained, and the dynamic evolution trend of the population within a future preset time period is obtained based on the real-time population density distribution map and historical population flow data. Potential bottleneck areas are identified based on the dynamic evolution trend of the population. A spatial attention heatmap is generated based on the potential bottleneck areas and the real-time crowd density distribution map, and high attention areas are obtained based on the spatial attention heatmap. The high-concern area is subjected to crowd movement field analysis to extract macroscopic fluid dynamic characteristics, and the presence of abnormal group behavior is determined based on the macroscopic fluid dynamic characteristics. When the aforementioned abnormal group behavior exists, a tiered early warning information is generated based on the abnormal group behavior.

[0006] Preferably, the step of obtaining the dynamic evolution trend of the population within a future preset time period based on the real-time population density distribution map and historical population flow data, and identifying potential bottleneck areas based on the dynamic evolution trend of the population, includes: Extract density features from the real-time crowd density distribution map and obtain historical flow patterns in historical crowd flow data that match the current timestamp; The density features and the historical flow patterns are input into a pre-trained spatiotemporal sequence prediction model, which outputs a predicted population density distribution map for a future preset time period. Calculate the crowd density gradient of each region in the predicted crowd density distribution map for each frame, and obtain the crowd flow direction field and velocity field based on the crowd density gradient; Based on the crowd flow direction field and velocity field, the path integral method is used to calculate the crowd pressure value of each area within a future preset time period. Areas where the stress value of the population exceeds a preset stress threshold are marked as potential bottleneck areas.

[0007] Preferably, the step of generating a spatial attention heatmap based on the potential bottleneck area and the real-time crowd density distribution map, and obtaining high-attention areas based on the spatial attention heatmap, includes: An initial area of ​​interest map is generated based on the spatial location information of the potential bottleneck area and the density values ​​in the real-time crowd density distribution map. The initial region of interest map is smoothed to generate a smooth region of interest map with a continuous weight distribution; The numerical range of the smoothed attention region map is normalized to a preset interval to form a basic spatial attention heatmap; Global action feature extraction is performed on the real-time video data to obtain a behavior feature map; High-attention regions are obtained based on the basic spatial attention heatmap and the behavioral feature map.

[0008] Preferably, the step of obtaining the high-attention region based on the basic spatial attention heatmap and the behavioral feature map includes: The behavior feature map is convolved to generate a behavior anomaly response map, wherein each pixel value in the behavior anomaly response map represents the probability of an abnormal behavior occurring at the corresponding location. The basic spatial attention heatmap and the behavioral anomaly response map are weighted and fused to generate a fused attention heatmap; The fused attention heatmap is binarized, and regions with pixel values ​​greater than a preset threshold are marked as candidate high attention regions; Morphological closing operations are performed on the candidate high-interest regions to form continuous high-interest regions; Based on the connected component analysis algorithm, the coordinates of the bounding rectangle of each continuous high-interest region are extracted, and the region defined by the coordinates of the bounding rectangle is determined as the high-interest region.

[0009] Preferably, the step of performing crowd movement field analysis on the high-concern area, extracting macroscopic fluid dynamics features, and determining whether there is abnormal group behavior based on the macroscopic fluid dynamics features includes: For the high-interest area, the crowd movement field is calculated using the optical flow method; Calculate the vortex field and divergence field of the crowd movement field; The average vorticity and average divergence of the high-interest region are calculated based on the vorticity field and the divergence field. Determine whether the average vorticity is greater than a preset vorticity threshold, and determine whether the average divergence is greater than a preset divergence threshold; If the average vorticity is greater than a preset vorticity threshold or the average divergence is greater than a preset divergence threshold, then it is determined that abnormal group behavior has occurred in the crowd.

[0010] Preferably, the step of generating tiered early warning information based on the abnormal group behavior when such abnormal group behavior exists includes: Obtain the anomaly type and spatial location of the aforementioned abnormal group behavior; The population anomaly intensity index is obtained based on the average vorticity and average divergence. Determine the warning interval in which the group anomaly intensity index is located, and obtain the corresponding warning level based on the warning interval; Based on the anomaly type, spatial location, and warning level, a graded early warning information is generated.

[0011] This invention also provides a system for monitoring crowd density and identifying abnormal behavior at large-scale exhibitions, comprising: The data acquisition module is used to acquire real-time video data collected by multiple monitoring devices in the exhibition venue, and generate a real-time crowd density distribution map by estimating the crowd density of the real-time video data. The bottleneck area identification module is used to acquire historical crowd flow data, obtain the dynamic evolution trend of the crowd within a future preset time period based on the real-time crowd density distribution map and historical crowd flow data, and identify potential bottleneck areas based on the dynamic evolution trend of the crowd. The attention module is used to generate a spatial attention heatmap based on the potential bottleneck area and the real-time crowd density distribution map, and to obtain high attention areas based on the spatial attention heatmap. The crowd anomaly detection module is used to perform crowd movement field analysis on the high-concern area, extract macroscopic fluid dynamics features, and determine whether there is any abnormal crowd behavior based on the macroscopic fluid dynamics features. The graded early warning module is used to generate graded early warning information based on the abnormal group behavior when such abnormal behavior exists.

[0012] Preferably, the attention module includes: The initial attention generation unit is used to generate an initial attention area map based on the spatial location information of the potential bottleneck area and the density value in the real-time crowd density distribution map. A smoothing processing unit is used to smooth the initial region of interest map to generate a smooth region of interest map with a continuous weight distribution. The normalized heatmap unit is used to normalize the numerical range of the smoothed region of interest map to a preset interval to form a basic spatial attention heatmap. An action feature extraction unit is used to extract global action features from the real-time video data to obtain a behavior feature map; The high-attention region acquisition unit is used to acquire high-attention regions based on the basic spatial attention heatmap and the behavioral feature map.

[0013] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, characterized in that the processor executes the computer program to implement the steps of a method for monitoring crowd density and identifying abnormal behavior at large-scale exhibitions.

[0014] The present invention also provides a computer-readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it implements the steps of a method for monitoring crowd density and identifying abnormal behavior at large-scale exhibitions.

[0015] The beneficial effects of this invention are as follows: This invention can achieve real-time control of the state of people in exhibition venues, advance risk prediction, and precise handling of anomalies. Specifically, it can identify potential bottleneck areas by combining historical crowd flow data to predict future crowd dynamic evolution trends, thus anticipating potential gathering risks and preventing delays in control measures. Furthermore, it can generate spatial attention heatmaps based on potential bottleneck areas and real-time crowd density distribution maps to identify high-attention areas, guiding the identification of abnormal behavior to focus on high-attention areas, improving the speed of monitoring and identification, and further improving the response efficiency of control measures. Moreover, by analyzing the crowd movement field in high-attention areas to extract macroscopic fluid dynamic characteristics to judge abnormal group behavior, it can prevent low identification accuracy when crowd density is high and mutual obstruction is severe, thereby improving the objectivity and accuracy of abnormal behavior judgment. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of a method flow according to an embodiment of this application.

[0017] Figure 2 This is a schematic diagram of the system structure according to an embodiment of this application.

[0018] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0019] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0020] like Figure 1 As shown, this application provides a method for monitoring crowd density and identifying abnormal behavior at large-scale exhibitions, including: S1, acquire real-time video data collected by multiple monitoring devices in the exhibition venue, and generate a real-time crowd density distribution map by estimating the crowd density of the real-time video data; S2, acquire historical crowd flow data, obtain the dynamic evolution trend of the crowd within a future preset time period based on the real-time crowd density distribution map and historical crowd flow data, and identify potential bottleneck areas based on the dynamic evolution trend of the crowd. S3, Generate a spatial attention heatmap based on the potential bottleneck area and the real-time crowd density distribution map, and obtain high attention areas based on the spatial attention heatmap; S4, perform crowd movement field analysis on the high-concern area, extract macroscopic fluid dynamics characteristics, and determine whether there is any abnormal group behavior based on the macroscopic fluid dynamics characteristics; S5, when the aforementioned abnormal group behavior exists, generate graded early warning information based on the aforementioned abnormal group behavior.

[0021] As described in steps S1-S5 above, the present invention can achieve real-time control of the status of people in exhibition venues, advance risk prediction and precise handling of anomalies, making control measures more timely and response more efficient in high-density crowd scenarios at large exhibitions.

[0022] In practical applications of large-scale exhibitions, venues typically feature vast spaces, fluctuating crowd density, and complex pedestrian flow paths. Furthermore, the high density of people during large-scale exhibitions means that failure to promptly monitor changes in crowd density and abnormal behavior can easily lead to safety accidents such as overcrowding and stampedes. Existing technologies primarily focus on monitoring individual behavioral characteristics and crowd density distribution. They identify anomalies by detecting rapid running or erratic collisions in video footage, or by identifying potential risks based on the overall macroscopic movement of the crowd. However, methods based on individual behavioral characteristics have low accuracy when crowd density is high and there is significant occlusion. Methods based on crowd density distribution detection tend to focus on static analysis of the current state. In contrast, crowd flow in large-scale exhibitions exhibits clear spatiotemporal correlations, making it difficult to predict future events simply by analyzing the current state. The potential for gathering risks can lead to delays in control measures. This invention, by combining historical crowd flow data to predict future crowd dynamics and identify potential bottleneck areas, can anticipate potential gathering risks and prevent delays in control measures. Furthermore, by generating spatial attention heatmaps based on potential bottleneck areas and real-time crowd density distribution maps, and identifying high-attention areas, it can guide abnormal behavior identification to focus on high-attention areas, improve the speed of monitoring and identification, and further enhance the responsiveness of control measures. Moreover, by analyzing crowd movement fields in high-attention areas to extract macroscopic fluid dynamic characteristics to judge abnormal group behavior, it can prevent low identification accuracy when crowd density is high and mutual occlusion is severe, thereby improving the objectivity and accuracy of abnormal behavior judgment and meeting the identification needs in high-density crowd scenarios of large-scale exhibitions.

[0023] In one embodiment of the present invention, the step of obtaining the dynamic evolution trend of the population within a future preset time period based on the real-time population density distribution map and historical population flow data, and identifying potential bottleneck areas based on the dynamic evolution trend of the population, includes: S21, extract the density features from the real-time crowd density distribution map, and obtain the historical flow pattern in the historical crowd flow data that matches the current timestamp; S22, Input the density features and the historical flow pattern into the pre-trained spatiotemporal sequence prediction model, and output the predicted population density distribution map for the future preset time period; S23, calculate the population density gradient of each region in the predicted population density distribution map of each frame, and obtain the population flow direction field and flow velocity field based on the population density gradient; S24, based on the crowd flow direction field and velocity field, the path integral method is used to calculate the crowd pressure value of each area within a preset time period in the future. The formula is: ; In the formula, This represents the crowd pressure value (used to quantify the degree of crowd congestion and potential congestion risk in a region; a higher value indicates that the region is more prone to bottlenecks). It represents the divergence operator (used to calculate the degree of divergence or convergence of a vector field, representing the difference between the outflow and inflow of people per unit volume). Indicates population density. It represents the velocity vector of the crowd (including the speed and direction of flow, representing the movement state of the crowd in space). Indicates a time variable (representing a future preset time period for calculating population stress values); S25, mark the areas where the population pressure value exceeds a preset pressure threshold as potential bottleneck areas.

[0024] Since the flow of people in large exhibition venues is not random and disorderly, but has obvious spatiotemporal correlation, as described in steps S21-S25 above, the present invention combines the current population density characteristics with historical flow patterns, and moves the risk identification node forward by predicting future population evolution trends, so as to achieve early identification of potential bottleneck areas.

[0025] The real-time crowd density distribution map is generated by processing the surveillance video data using a CNN model based on a ResNet-50 encoder-decoder structure. The historical crowd flow data comes from historical surveillance videos and gate statistics of similar exhibitions. This data is stored in the venue's MySQL database according to the categories of "exhibition date-time period-area". When obtaining historical flow patterns that match the current timestamp, historical data from the same period of the exhibition in recent years are retrieved using SQL statements. The average density change rate of each grid area within that time period is calculated to form a historical flow pattern feature vector consistent with the real-time density feature dimension.

[0026] In the step of inputting the density features and historical flow patterns into a pre-trained spatiotemporal sequence prediction model and outputting a predicted crowd density distribution map for a future preset time period, the spatiotemporal sequence prediction model adopts an LSTM-based network structure. The input layer receives a concatenated vector of real-time density features and historical flow pattern features, and the dimensionality is compressed through an embedding layer to reduce redundant information. The intermediate layer sets two LSTM units, each with 512 hidden units, and introduces a dropout coefficient of 0.3 to prevent model overfitting and ensure generalization ability under different exhibition scenarios. Subsequently, the feature dimension is restored to the original dimension through a fully connected layer. Finally, the output layer amplifies the average grid density value corresponding to the vector to the target pixel through bilinear interpolation to generate a predicted crowd density distribution map for a future preset time period.

[0027] Based on the predicted crowd density distribution map, the crowd density gradient of each region in each frame of the predicted crowd density distribution map is calculated, and the crowd flow direction field and velocity field are obtained according to the crowd density gradient. The density gradient is a vector whose direction points to the direction of the fastest increase in density, and its magnitude represents the degree of drastic change in density. The positive density gradient direction, that is, the direction of the fastest increase in density, is the potential flow direction of the crowd. Based on this, the present invention further derives the crowd flow direction field and velocity field from the density gradient field. Compared with the existing technical solutions that statistically analyze the density or flow of the crowd at the current moment, this solution can predict the potential clustering risk in the next few minutes earlier and improve the timeliness of safety management measures.

[0028] Subsequently, based on the direction and velocity fields of crowd flow, the path integral method is used to calculate the crowd pressure value of each area within a preset time period, enabling a deeper risk assessment. Here, the crowd pressure value quantifies the internal interaction forces generated by crowd convergence and flow obstruction, representing the net inflow or outflow of crowds per unit area. It is a consistently positive divergence integral value, indicating that the area is continuously accumulating crowds and the pressure is continuously increasing. Finally, areas where the crowd pressure value exceeds the preset pressure threshold based on the venue structure and safety standards are marked as potential bottleneck areas, thereby improving the early warning capability in complex scenarios of large-scale exhibitions.

[0029] In one embodiment of the present invention, the step of generating a spatial attention heatmap based on the potential bottleneck area and the real-time crowd density distribution map, and obtaining high-attention areas based on the spatial attention heatmap, includes: S31, an initial attention area map is generated based on the spatial location information of the potential bottleneck area and the density value in the real-time crowd density distribution map; S32, the initial region of interest map is smoothed to generate a smooth region of interest map with continuous weight distribution; S33, normalize the numerical range of the smoothed attention region map to a preset interval to form a basic spatial attention heatmap; S34, global action feature extraction is performed on the real-time video data to obtain a behavior feature map; S35, obtain the high attention region based on the basic spatial attention heatmap and the behavioral feature map.

[0030] As described in steps S31-S35 above, since the spatial distribution of crowd safety risks is not uniform, the risk is the result of the combined effect of multiple factors. It is not only related to the current crowd density, but also to whether the physical spatial structure of the crowd is prone to forming bottlenecks, and is closely related to the current behavioral dynamics of the crowd in the area. Therefore, this solution uses a spatial attention mechanism to dynamically focus limited computing and analysis resources on high-risk areas where anomalies are most likely to occur, thereby achieving in-depth and efficient analysis of key areas while ensuring full-field monitoring coverage, and improving the speed of monitoring and analysis.

[0031] First, this solution generates an initial area of ​​concern map based on the spatial location information of potential bottleneck areas and the density values ​​in the real-time crowd density distribution map. The initial area of ​​concern map can assign higher weights to potential bottleneck areas to highlight their risk priority. Specifically, for pixels in the real-time crowd density distribution map that are within the bounding rectangle of the bottleneck area, their density values ​​are multiplied by a higher weighting coefficient; for pixels in non-bottleneck areas, their original density values ​​are kept unchanged, thus forming the initial area of ​​concern map.

[0032] Next, the initial region of interest map is smoothed to generate a smooth region of interest map with a continuous weight distribution. This is because the initial map generated directly from the data and coordinates often has sharp edges and abrupt numerical changes, which does not match the diffusion and continuity of risks in the real world. Therefore, by applying a Gaussian filter for convolution operations, these discontinuous abrupt changes can be eliminated, generating a smooth region of interest map with a continuous weight distribution, so that the influence of high-risk areas can naturally transition to their surroundings.

[0033] Then, the numerical range of the smooth attention region map is normalized to a preset interval, which is between 0 and 1, thereby forming a basic spatial attention heatmap. The basic spatial attention heatmap can transform the previous comprehensive risk representation into a standard weight value, so that the value of each pixel clearly represents the relative attention priority of that spatial location within the range of 0-1.

[0034] On another parallel path, the system performs global motion feature extraction on real-time video data to obtain behavioral feature maps. Here, a 3D CNN model with a C3D network structure is used for global motion feature extraction to capture spatiotemporal motion information in video sequences, adapting to the dynamic characteristics of crowd behavior. Specifically, the 3D CNN model is input with real-time video data, selecting 16 consecutive video frames, with the input size adjusted to 640×480×16×3 (height×width×number of frames×number of channels). The model structure includes 5 layers of 3D convolutional layers (all convolutional kernels are 3×3×3, stride is 1×1×1, and padding is set to 1×1×1 to ensure that the output size is consistent with the input), 4 layers of max pooling layers (pooling kernels are 2×2×2, stride is 2×2×2, used to compress feature dimensions), and 2 layers of fully connected layers (the number of neurons is 4096 and 2048 respectively). The final output is a 16×12×16 action feature map (256 channels). Each feature pixel corresponds to the action feature intensity of a 32×32 pixel region in the original video (the higher the value, the greater the probability of abnormal action in that region). To adapt to the subsequent size matching with the basic spatial attention heatmap, the 256-channel action feature map is globally averaged and pooled to transform it into a single-channel feature map. Then, it is enlarged to 640×480 pixels using a bilinear interpolation algorithm to finally obtain an action feature map with the same size as the basic spatial attention heatmap.

[0035] Finally, high-attention areas are obtained based on the basic spatial attention heatmap and behavioral feature map. By fusing multiple features and comprehensively considering crowd density, bottleneck risk and behavioral risk, the accuracy of high-attention area positioning is improved, thereby enhancing the efficiency and effectiveness of security monitoring in high-density crowd scenarios at large-scale exhibitions.

[0036] In one embodiment of the present invention, the step of obtaining the high-attention region based on the basic spatial attention heatmap and the behavioral feature map includes: S351, perform convolution processing on the behavior feature map to generate a behavior anomaly response map, wherein each pixel value in the behavior anomaly response map represents the probability of an abnormal behavior occurring at the corresponding location. S352, The basic spatial attention heatmap and the behavioral anomaly response map are weighted and fused to generate a fused attention heatmap; S353, perform binarization processing on the fused attention heatmap, and mark the regions with pixel values ​​greater than a preset threshold as candidate high attention regions; S354, perform morphological closing operation on the candidate high-interest regions to form continuous high-interest regions; S355, Based on the connected component analysis algorithm, extract the coordinates of the bounding rectangle of each continuous high-interest region, and determine the region defined by the coordinates of the bounding rectangle as the high-interest region.

[0037] As described in steps S351-S355 above, the present invention generates an abnormal behavior response map by performing convolution processing on the behavior feature map, and then weightedly fuses it with the basic spatial attention heatmap. After binarization, morphological closing operation and connected component analysis, the high-attention areas that truly need to be monitored are accurately screened from a large range of venue areas, further improving the efficiency and accuracy of anomaly identification.

[0038] Specifically, the first step is to perform convolution processing on the behavior feature map to generate a behavior anomaly response map. Convolution processing can further enhance the abnormal signal. In the resulting behavior anomaly response map, each pixel value represents the probability of abnormal behavior occurring at the corresponding location. For example, the convolution processing uses a combination of 1×1 and 3×3 convolution kernels. The 1×1 convolution kernels (64 in number) are used to increase the feature dimension and enhance the model's ability to express behavior features. The 3×3 convolution kernels (32 in number, stride 1, padding=1) are used to capture the local spatial correlation of behavior features. The activation function used in the convolution processing is ReLU to avoid the gradient vanishing problem. Finally, the output feature map is normalized to the [0,1] interval by the Sigmoid function to generate the behavior anomaly response map. The closer the pixel value is to 1, the higher the probability of abnormal behavior occurring at that location.

[0039] The second step is to perform a weighted fusion of the basic spatial attention heatmap and the behavioral anomaly response map to generate a fused attention heatmap. The basic spatial attention heatmap focuses on reflecting the spatial risk of a region, while the behavioral anomaly response map focuses on reflecting the behavioral risk of a region. The fusion of these two types of features can more comprehensively measure the regional risk. The fusion weights are set based on the scenario requirements of large-scale exhibitions. Spatial risk is the foundation. If a region has no spatial risk (low density, non-bottleneck), even if there are individual abnormal behaviors, the overall risk is low. Therefore, the weight of the basic spatial attention heatmap is set to 0.6. Behavioral risk is supplementary. It can promptly identify regions with low spatial risk but abnormal behavior. Therefore, the weight of the behavioral anomaly response map is set to 0.4, and the sum of the weights is 1 to ensure the rationality of the fusion result. Weighted fusion employs a pixel-by-pixel multiplication and summation method. For example, the base spatial attention value for a potential bottleneck region is 0.7, and the behavioral anomaly response value is 0.3, resulting in a fused value of 0.6 × 0.7 + 0.4 × 0.3 = 0.54. Similarly, for a non-bottleneck region, the base spatial attention value is 0.3, and the behavioral anomaly response value is 0.9, resulting in a fused value of 0.6 × 0.3 + 0.4 × 0.9 = 0.54. This fusion method allows for reasonable risk quantification of regions with different risk types, avoiding misjudgments caused by single features and thus improving the accuracy of anomaly identification.

[0040] The third step is to binarize the fused attention heatmap, marking areas with pixel values ​​greater than a preset threshold as candidate high-attention areas. The numerical range of the fused attention heatmap is [0,1]. High-risk and low-risk areas are clearly distinguished by the preset threshold. The preset threshold is determined based on historical abnormal event data statistics. For example, analyzing the numerical distribution of the fused attention heatmap when 500 abnormal events occurred in exhibitions over the past 3 years, it was found that more than 90% of the abnormal events occurred in areas with a fusion value greater than 0.62. Therefore, the preset threshold is set to 0.62. This preset threshold can ensure the coverage of high-risk areas while effectively eliminating interference from low-risk areas. The binarization process uses a fixed threshold method. If a pixel value in the fused attention heatmap is greater than 0.62, it is marked as 1 (high risk); otherwise, it is marked as 0 (low risk). Finally, a binarized image is generated, and the area formed by the pixels marked as 1 is the candidate high-attention area.

[0041] The fourth step is to perform morphological closing operations on the candidate high-interest regions to form continuous high-interest regions. Due to factors such as image noise and local occlusion, the candidate high-interest regions may contain internal holes (such as local 0-value regions caused by the shadow of the display stand) or discrete small regions (such as 1-value points of a single pixel). These incomplete regions will affect the accuracy of subsequent anomaly identification. The scattered candidate regions are connected to form a more complete and continuous high-interest region that is more in line with the distribution of real people.

[0042] The fifth step is to extract the bounding rectangle coordinates of each continuous region of interest (GI) based on connected component analysis (CBI) algorithms, and then define the region bounded by the bounding rectangle coordinates as the GI. The purpose of CBI is to extract complete continuous regions from a binarized image and simplify the region representation using bounding rectangles. For example, an 8-neighborhood-based CBI algorithm is used. By scanning the binarized image, all connected components with a pixel value of 1 are marked, and each connected component is assigned a unique identifier. Then, the minimum bounding rectangle of each connected component is calculated—the smallest rectangle containing all pixels of that connected component. The coordinates of the top-left and bottom-right corners of the rectangle are extracted. Simultaneously, to exclude invalid small regions, an area threshold is set. If the area of ​​a connected component is less than the threshold, it is considered a noise region and is removed. If the area is greater than or equal to the threshold, the region bounded by its bounding rectangle coordinates is defined as a GI. Subsequent abnormal behavior identification only needs to analyze the GI, significantly reducing the computational scope and thus improving the efficiency of abnormal behavior identification in large-scale exhibition environments.

[0043] In one embodiment of the present invention, the step of performing crowd movement field analysis on the high-concern area, extracting macroscopic fluid dynamic features, and determining whether there is abnormal group behavior based on the macroscopic fluid dynamic features includes: S41, For the high-interest area, the crowd movement field is calculated using the optical flow method; S42, Calculate the vortex field and divergence field of the crowd movement field; The formula for calculating vorticity is: ; In the formula, Vorticity (used to quantify the intensity and direction of rotation in a crowd's movement field; positive values ​​indicate counterclockwise rotation, negative values ​​indicate clockwise rotation, and the larger the absolute value, the more intense the rotation). Represents the curl operator, Represents the velocity vector of the crowd; The formula for calculating divergence is: ; In the formula, This represents the divergence (used to quantify the degree of dispersion or convergence of a crowd's movement field; a positive value indicates that the crowd disperses in all directions, such as in a panicked escape, while a negative value indicates that the crowd converges towards the center; the larger the absolute value, the more obvious the dispersion or convergence). Denotes the divergence operator, Represents the velocity vector of the crowd; S43, Calculate the average vorticity and average divergence of the high-interest region based on the vorticity field and the divergence field; S44, determine whether the average vorticity is greater than a preset vorticity threshold, and determine whether the average divergence is greater than a preset divergence threshold; S45, if the average vorticity is greater than a preset vorticity threshold or the average divergence is greater than a preset divergence threshold, then it is determined that abnormal group behavior has occurred in the crowd.

[0044] As described in steps S41-S45 above, when the crowd density reaches a certain level, the individual's freedom of movement is greatly restricted, and their behavior will inevitably be affected by the overall flow. At this time, analysis from the perspective of continuous medium is more effective than tracking discrete individuals. Therefore, this invention introduces fluid dynamics analysis methods to quantify and characterize the macroscopic movement patterns of the crowd in the high-interest area and make anomaly judgments, thereby effectively identifying abnormal group behaviors that are difficult for individual behavior detection methods to capture in high-density occlusion scenarios.

[0045] Based on the characteristics of large-scale exhibitions and the physical laws of crowd movement, the core feature of abnormal group behavior is not a simple summation of individual behaviors, but a significant anomaly in the overall movement pattern of the crowd. For example, during a panic swirl, a local rotating flow field is formed in the crowd (due to information asymmetry and misjudgment, individuals cannot accurately determine the source of danger and react solely based on the behavior of those around them; seeing others run, they follow suit, leading to a herd mentality. When space is limited and exits are blocked, in a high-density environment, once someone moves in a certain direction, others are forced to follow, forming a chain reaction. If exits are blocked or guidance is inadequate, the crowd will fall into a cyclical movement. Panic is contagious, and emotions spread rapidly, causing the group to lose its ability to make rational judgments, resulting in blind following and reverse pushing). When a group moves at high speed in a narrow space, it is easy to form a local rotating flow field due to friction, collision and inertia. When the group scatters, the flow field characteristics of the crowd will be rapidly diverging in all directions (the originally relatively stable crowd suddenly moves rapidly in all directions with a certain point as the center). This kind of macroscopic movement pattern cannot be captured by single individual behavior detection. At the same time, in large exhibitions, the high density of crowds will seriously block each other, and individual behavior detection will easily fail to extract features due to target occlusion. This invention obtains the crowd movement field by optical flow method and extracts macroscopic features such as vorticity and divergence to realize the judgment of different types of abnormal group behavior. It is not only suitable for high-density occlusion scenarios, but also broadens the coverage of anomaly recognition and makes up for the shortcomings of existing technologies.

[0046] Specifically, the first step is to calculate the crowd movement field for high-interest areas using optical flow. The Farneback optical flow algorithm is selected, which generates a dense optical flow field to meet the requirement that "every pixel in the crowd movement field needs motion information." Compared to sparse optical flow algorithms, this is more suitable for macroscopic flow field analysis scenarios. The algorithm parameters are calibrated based on the characteristics of large-scale exhibition monitoring videos; for example, the number of pyramid layers is set to 3, the window size to 15×15, the number of iterations to 5, and the polynomial expansion window size to 5×5. During calculation, two consecutive frames of real-time video data corresponding to the high-interest area are input (the frame interval is the reciprocal of the monitoring device's frame rate, such as 40ms for a frame rate of 25fps). The output is the optical flow vector of each pixel in that area. To convert pixel displacement into actual motion velocity, the physical parameters of the monitoring device are also used for conversion, ultimately yielding the crowd movement field containing the actual motion velocity and direction of each pixel.

[0047] Then, the vorticity and divergence fields of the crowd movement field are calculated. Vorticity and divergence are the core macroscopic features describing fluid motion. Analogous to crowd movement, vorticity is used to quantify the rotation intensity and direction of the crowd, while divergence is used to quantify the degree of dispersion or convergence of the crowd. The combination of the two can comprehensively characterize the abnormal movement patterns of the group. Positive vorticity values ​​represent counterclockwise rotation of the crowd, while negative values ​​represent clockwise rotation. The larger the absolute value, the higher the rotation intensity. Positive divergence values ​​represent the crowd dispersing in all directions, while negative values ​​represent the crowd converging towards the center. The larger the absolute value, the more significant the degree of dispersion or convergence.

[0048] Next, the average vorticity and average divergence of the high-interest area are statistically analyzed based on the vorticity and divergence fields. Since the vorticity or divergence values ​​of a single pixel may be affected by noise, directly using them for judgment can easily lead to misjudgments. Therefore, this invention obtains macroscopic features through regional statistics. When calculating the average vorticity, the absolute values ​​of the vorticity values ​​of all pixels in the vorticity field are first taken (to eliminate the interference of rotation direction on anomaly judgment, retaining only intensity information), and then the arithmetic mean of the absolute values ​​of all pixels is calculated. When calculating the average divergence, the arithmetic mean of the divergence values ​​of all pixels in the divergence field is directly calculated. These macroscopic statistical values ​​can effectively filter noise and reflect the overall degree of motion anomalies in the region.

[0049] Next, it is determined whether the average vorticity exceeds a preset vorticity threshold and whether the average divergence exceeds a preset divergence threshold. The preset thresholds are determined based on a comparative analysis of historical abnormal event data from large-scale exhibitions and normal crowd movement data. This ensures that the thresholds effectively distinguish between abnormal and normal states without causing false alarms due to oversensitivity. For example, by statistically analyzing the average vorticity and average divergence during abnormal events such as panic spinning and mass scattering at exhibitions over the past three years, and comparing them with the corresponding values ​​during normal visiting periods, it was found that the average vorticity of abnormal events is generally greater than 0.8 rad / s, and the average divergence is generally greater than 0.5 1 / s, while the average vorticity during normal periods is mostly less than 0.3 rad / s, and the average divergence is mostly between -0.2 and 0.2 1 / s. Therefore, the preset vorticity threshold is set to 0.8 rad / s, and the preset divergence threshold is set to 0.5 1 / s. The judgment logic uses "OR logic," meaning that as long as either the average vorticity exceeds the preset vorticity threshold or the average divergence exceeds the preset divergence threshold, it is determined that there is abnormal group behavior; both conditions do not need to be met simultaneously to determine an anomaly. If both the average vorticity and average divergence are less than the corresponding thresholds, it is determined that there is no abnormal group behavior, and the system continues to monitor subsequent frames.

[0050] Finally, if the average vorticity is greater than a preset vorticity threshold or the average divergence is greater than a preset divergence threshold, it is determined that there is abnormal group behavior in the crowd. During the determination process, the key basis for the abnormality determination is recorded simultaneously, including the abnormality type, the coordinates of the high-concern area where the abnormality occurred, and the specific values ​​of the average vorticity and average divergence. If it is determined that there is no abnormality, the motion feature data of the current high-concern area is stored in the historical database. Through this invention, potential group risks can still be identified stably and objectively in the core area of ​​the exhibition where the crowd density is high and the occlusion between individuals is severe.

[0051] In one embodiment of the present invention, the step of generating graded early warning information based on the abnormal group behavior when the abnormal group behavior exists includes: S51, Obtain the abnormal type and spatial location of the abnormal group behavior; S52, the population anomaly intensity index is obtained based on the average vorticity and average divergence, using the following formula: ; In the formula, The group anomaly intensity index (used to comprehensively quantify the severity of abnormal group behavior; the higher the value, the more severe the anomaly). This represents the vorticity weighting coefficient (used to adjust the importance of vorticity in anomaly intensity assessment, determined based on historical anomalous behavior samples, analyzing the frequency and severity of accidents caused by past panic-induced spinning behavior; if such anomalies are more dangerous, then...). The value is biased towards 0.6, and vice versa, towards 0.4. The default value is set to 0.5. Indicates the average vorticity. This indicates the preset vorticity threshold. This represents the divergence weighting coefficient (used to adjust the importance of divergence in anomaly strength assessment). Indicates the average divergence. Indicates the preset divergence threshold; S53, determine the warning interval where the group abnormality intensity index is located, and obtain the corresponding warning level according to the warning interval; S54, Generate graded early warning information based on the anomaly type, spatial location, and early warning level.

[0052] As described in steps S51-S54 above, the present invention establishes a graded early warning mechanism, which transforms the abnormal group behaviors identified in the previous steps into graded early warning information with clear guiding significance according to their severity, type and location information, thereby realizing the precise scheduling of security resources and the differentiated implementation of emergency response measures.

[0053] Specifically, the scheme first obtains the anomaly type and spatial location of the aforementioned abnormal group behavior. The anomaly type and spatial location are directly derived from the determination result of claim 5 and the high-concern area upon which it is based. The anomaly type is mainly distinguished into panic-driven spinning behavior dominated by high vorticity and panic-driven scattering behavior dominated by high divergence. The spatial location is determined by the coordinates of the bounding rectangle of the high-concern area finally determined in claim 4, which clearly defines the physical scope of the abnormal behavior.

[0054] Subsequently, the system obtains the population anomaly intensity index based on the average vorticity and average divergence. The population anomaly intensity index is integrated into a comprehensive risk scalar through a weighted formula, where the specific values ​​of the weight coefficients represent the hazard assessment of different types of abnormal behavior.

[0055] Based on this, the warning interval of the abnormal intensity index of the group is determined, and the corresponding warning level is obtained according to the warning interval. The division of the warning interval is pre-set based on historical event data analysis and safety plans.

[0056] Finally, based on the anomaly type, spatial location, and warning level, the final graded warning information is generated. This information is assembled into structured data objects and pushed to the security command center through a human-computer interaction interface. For example: First, the warning level is clearly marked as "Level 1 Warning," "Level 2 Warning," or "Level 3 Warning," and distinguished by color (blue for Level 1, yellow for Level 2, and red for Level 3) to improve visual recognition; second, anomaly details include the anomaly type and spatial location, such as "Anomaly type: Rotational anomaly; Location: Main exhibition hall, Area A"; third, the anomaly type provides a reference for security personnel; fourth, it may also include handling suggestions, formulating targeted guidance based on the warning level and anomaly type. These suggestions are generated based on a historical handling plan database, such as: Level 1 warning suggestion "Dispatch 2 regional security personnel to the scene to guide the crowd to move in an orderly manner using handheld megaphones"; Level 2 warning suggestion "Dispatch 4-6..." "Support with mobile security personnel, close secondary entrances around the abnormal area, and activate emergency broadcasts for targeted alerts"; the Level 3 warning recommends "activating the venue's emergency response plan, dispatching all nearby security personnel to the scene, coordinating with medical and fire departments to stand by, closing all entrances to the abnormal area, and guiding surrounding crowds to evacuate to safe areas." In this way, a balance was achieved between maximizing response efficiency and minimizing on-site interference while ensuring the safety of large-scale exhibitions, thus improving the professional level of safety management.

[0057] A method for monitoring crowd density and identifying abnormal behavior at large-scale exhibitions includes: The system acquires real-time video data collected by multiple monitoring devices within the exhibition venue, and generates a real-time crowd density distribution map by estimating the crowd density of the real-time video data. Historical population flow data is obtained, and the dynamic evolution trend of the population within a future preset time period is obtained based on the real-time population density distribution map and historical population flow data. Potential bottleneck areas are identified based on the dynamic evolution trend of the population. A spatial attention heatmap is generated based on the potential bottleneck areas and the real-time crowd density distribution map, and high attention areas are obtained based on the spatial attention heatmap. The high-concern area is subjected to crowd movement field analysis to extract macroscopic fluid dynamic characteristics, and the presence of abnormal group behavior is determined based on the macroscopic fluid dynamic characteristics. When the aforementioned abnormal group behavior exists, a tiered early warning information is generated based on the abnormal group behavior.

[0058] Preferably, the step of obtaining the dynamic evolution trend of the population within a future preset time period based on the real-time population density distribution map and historical population flow data, and identifying potential bottleneck areas based on the dynamic evolution trend of the population, includes: Extract density features from the real-time crowd density distribution map and obtain historical flow patterns in historical crowd flow data that match the current timestamp; The density features and the historical flow patterns are input into a pre-trained spatiotemporal sequence prediction model, which outputs a predicted population density distribution map for a future preset time period. Calculate the crowd density gradient of each region in the predicted crowd density distribution map for each frame, and obtain the crowd flow direction field and velocity field based on the crowd density gradient; Based on the crowd flow direction field and velocity field, the path integral method is used to calculate the crowd pressure value of each area within a future preset time period. Areas where the stress value of the population exceeds a preset stress threshold are marked as potential bottleneck areas.

[0059] Preferably, the step of generating a spatial attention heatmap based on the potential bottleneck area and the real-time crowd density distribution map, and obtaining high-attention areas based on the spatial attention heatmap, includes: An initial area of ​​interest map is generated based on the spatial location information of the potential bottleneck area and the density values ​​in the real-time crowd density distribution map. The initial region of interest map is smoothed to generate a smooth region of interest map with a continuous weight distribution; The numerical range of the smoothed attention region map is normalized to a preset interval to form a basic spatial attention heatmap; Global action feature extraction is performed on the real-time video data to obtain a behavior feature map; High-attention regions are obtained based on the basic spatial attention heatmap and the behavioral feature map.

[0060] Preferably, the step of obtaining the high-attention region based on the basic spatial attention heatmap and the behavioral feature map includes: The behavior feature map is convolved to generate a behavior anomaly response map, wherein each pixel value in the behavior anomaly response map represents the probability of an abnormal behavior occurring at the corresponding location. The basic spatial attention heatmap and the behavioral anomaly response map are weighted and fused to generate a fused attention heatmap; The fused attention heatmap is binarized, and regions with pixel values ​​greater than a preset threshold are marked as candidate high attention regions; Morphological closing operations are performed on the candidate high-interest regions to form continuous high-interest regions; Based on the connected component analysis algorithm, the coordinates of the bounding rectangle of each continuous high-interest region are extracted, and the region defined by the coordinates of the bounding rectangle is determined as the high-interest region.

[0061] Preferably, the step of performing crowd movement field analysis on the high-concern area, extracting macroscopic fluid dynamics features, and determining whether there is abnormal group behavior based on the macroscopic fluid dynamics features includes: For the high-interest area, the crowd movement field is calculated using the optical flow method; Calculate the vortex field and divergence field of the crowd movement field; The average vorticity and average divergence of the high-interest region are calculated based on the vorticity field and the divergence field. Determine whether the average vorticity is greater than a preset vorticity threshold, and determine whether the average divergence is greater than a preset divergence threshold; If the average vorticity is greater than a preset vorticity threshold or the average divergence is greater than a preset divergence threshold, then it is determined that abnormal group behavior has occurred in the crowd.

[0062] Preferably, the step of generating tiered early warning information based on the abnormal group behavior when such abnormal group behavior exists includes: Obtain the anomaly type and spatial location of the aforementioned abnormal group behavior; The population anomaly intensity index is obtained based on the average vorticity and average divergence. Determine the warning interval in which the group anomaly intensity index is located, and obtain the corresponding warning level based on the warning interval; Based on the anomaly type, spatial location, and warning level, a graded early warning information is generated.

[0063] like Figure 2 As shown, the present invention also provides a large-scale exhibition crowd density monitoring and abnormal behavior identification system, comprising: The data acquisition module is used to acquire real-time video data collected by multiple monitoring devices in the exhibition venue, and generate a real-time crowd density distribution map by estimating the crowd density of the real-time video data. The bottleneck area identification module is used to acquire historical crowd flow data, obtain the dynamic evolution trend of the crowd within a future preset time period based on the real-time crowd density distribution map and historical crowd flow data, and identify potential bottleneck areas based on the dynamic evolution trend of the crowd. The attention module is used to generate a spatial attention heatmap based on the potential bottleneck area and the real-time crowd density distribution map, and to obtain high attention areas based on the spatial attention heatmap. The crowd anomaly detection module is used to perform crowd movement field analysis on the high-concern area, extract macroscopic fluid dynamics features, and determine whether there is any abnormal crowd behavior based on the macroscopic fluid dynamics features. The graded early warning module is used to generate graded early warning information based on the abnormal group behavior when such abnormal behavior exists.

[0064] Preferably, the attention module includes: The initial attention generation unit is used to generate an initial attention area map based on the spatial location information of the potential bottleneck area and the density value in the real-time crowd density distribution map. A smoothing processing unit is used to smooth the initial region of interest map to generate a smooth region of interest map with a continuous weight distribution. The normalized heatmap unit is used to normalize the numerical range of the smoothed region of interest map to a preset interval to form a basic spatial attention heatmap. An action feature extraction unit is used to extract global action features from the real-time video data to obtain a behavior feature map; The high-attention region acquisition unit is used to acquire high-attention regions based on the basic spatial attention heatmap and the behavioral feature map.

[0065] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, characterized in that the processor executes the computer program to implement the steps of a method for monitoring crowd density and identifying abnormal behavior at large-scale exhibitions.

[0066] The present invention also provides a computer-readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it implements the steps of a method for monitoring crowd density and identifying abnormal behavior at large-scale exhibitions.

[0067] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, apparatus, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

[0068] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A large-scale exhibition crowd density monitoring and abnormal behavior identification method, characterized in that, The method comprises the following steps: acquiring real-time video data collected by a plurality of monitoring devices in an exhibition venue, and generating a real-time crowd density distribution map by performing crowd density estimation on the real-time video data; acquiring historical crowd flow data, obtaining a crowd dynamic evolution trend in a future preset time period according to the real-time crowd density distribution map and the historical crowd flow data, and identifying a potential bottleneck area based on the crowd dynamic evolution trend; generating a spatial attention heat map according to the potential bottleneck area and the real-time crowd density distribution map, and obtaining a high attention area based on the spatial attention heat map; performing crowd motion field analysis on the high attention area, extracting macro fluid dynamics characteristics, and determining whether there is a group abnormal behavior according to the macro fluid dynamics characteristics; when the group abnormal behavior exists, generating a hierarchical warning information according to the group abnormal behavior. 2.The method of claim 1, wherein, The step of obtaining a crowd dynamic evolution trend in a future preset time period according to the real-time crowd density distribution map and the historical crowd flow data, and identifying a potential bottleneck area based on the crowd dynamic evolution trend comprises: extracting density features in the real-time crowd density distribution map, and acquiring a historical flow pattern matching a current time stamp in the historical crowd flow data; inputting the density features and the historical flow pattern into a pre-trained spatio-temporal sequence prediction model to output a predicted crowd density distribution map in a future preset time period; calculating a crowd density gradient of each region in each frame of the predicted crowd density distribution map, and obtaining a crowd flow direction field and a flow velocity field according to the crowd density gradient; calculating a crowd pressure value of each region in a future preset time period by using a path integral method according to the crowd flow direction field and the flow velocity field; marking a region with a crowd pressure value exceeding a preset pressure threshold as a potential bottleneck area.

3. The method according to claim 1, wherein, The step of generating a spatial attention heat map according to the potential bottleneck area and the real-time crowd density distribution map, and obtaining a high attention area based on the spatial attention heat map comprises: generating an initial attention area map based on spatial position information of the potential bottleneck area and density values in the real-time crowd density distribution map; performing smoothing processing on the initial attention area map to generate a smoothed attention area map with continuous weight distribution; normalizing a value range of the smoothed attention area map to a preset interval to form a basic spatial attention heat map; performing global action feature extraction on the real-time video data to obtain a behavior feature map; obtaining a high attention area according to the basic spatial attention heat map and the behavior feature map.

4. The method according to claim 3, wherein, The step of obtaining a high attention area according to the basic spatial attention heat map and the behavior feature map comprises: performing convolution processing on the behavior feature map to generate a behavior anomaly response map, wherein each pixel value in the behavior anomaly response map represents a probability of abnormal behavior occurring at a corresponding position; performing weighted fusion of the basic spatial attention heat map and the behavior anomaly response map to generate a fusion attention heat map; performing binaryzation processing on the fusion attention heat map, and marking a region with a pixel value greater than a preset threshold as a candidate high attention area; Performing morphological closing operation on the candidate high attention region to form a continuous high attention region; Extracting the circumscribed rectangle coordinates of each continuous high attention region based on a connected domain analysis algorithm, and determining the region bounded by the circumscribed rectangle coordinates as a high attention region.

5. The method according to claim 1, wherein, The step of performing crowd motion field analysis on the high attention region, extracting macro fluid dynamics characteristics, and determining whether there is a group abnormal behavior according to the macro fluid dynamics characteristics comprises: Calculating the crowd motion field by optical flow method for the high attention region; Calculating the vorticity field and divergence field of the crowd motion field; According to the vorticity field and the divergence field, the average vorticity and the average divergence of the high attention region are calculated; Determine whether the average vorticity is greater than a preset vorticity threshold, and determine whether the average divergence is greater than a preset divergence threshold; If the average vorticity is greater than the preset vorticity threshold or the average divergence is greater than the preset divergence threshold, it is determined that there is a group abnormal behavior in the crowd.

6. The method according to claim 5, wherein, The step of generating a hierarchical warning information according to the group abnormal behavior when the group abnormal behavior exists comprises: Obtaining the abnormal type and spatial position of the group abnormal behavior; Obtaining the group abnormal intensity index according to the average vorticity and average divergence; Determine the warning interval of the group abnormal intensity index, and obtain the corresponding warning level according to the warning interval; Generate hierarchical warning information according to the abnormal type, spatial position and warning level.

7. A large-scale exhibition crowd density monitoring and abnormal behavior identification system, characterized in that, Comprise: Data acquisition module, for acquiring real-time video data collected by multiple monitoring devices in the exhibition venue, and generating real-time crowd density distribution map by crowd density estimation on the real-time video data; Bottleneck area identification module, for obtaining historical crowd flow data, obtaining crowd dynamic evolution trend in a future preset time period according to the real-time crowd density distribution map and historical crowd flow data, and identifying and obtaining potential bottleneck area based on the crowd dynamic evolution trend; Attention module, for generating spatial attention heat map according to the potential bottleneck area and real-time crowd density distribution map, and obtaining high attention region based on the spatial attention heat map; Group anomaly detection module, for performing crowd motion field analysis on the high attention region, extracting macro fluid dynamics characteristics, and determining whether there is a group abnormal behavior according to the macro fluid dynamics characteristics; Hierarchical warning module, for generating hierarchical warning information according to the group abnormal behavior when the group abnormal behavior exists.

8. The large-scale convention crowd density monitoring and abnormal behavior identification system according to claim 7, wherein, The attention module comprises: Initial attention generation unit, for generating initial attention region map based on the spatial position information of the potential bottleneck area and the density value in the real-time crowd density distribution map; Smoothing processing unit, for smoothing the initial attention region map to generate a smooth attention region map with continuous weight distribution; Normalization heat map unit, for normalizing the value range of the smooth attention region map to a preset interval to form a basic spatial attention heat map; Action feature extraction unit, for performing global action feature extraction on the real-time video data to obtain behavior feature map; A high attention region acquisition unit is configured to acquire a high attention region according to the basic spatial attention heat map and the behavior feature map. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. The processor implements the steps of the method of any one of claims 1 to 6 when executing the computer program.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program, when executed by the processor, implements the steps of the method of any one of claims 1 to 6. The computer program, when executed by the processor, implements the steps of the method of any one of claims 1 to 6.

Images